The Dockerfile layer that builds flash attention uses the PYTORCH_ROCM_ARCH argument. But since we don't have ARG PYTORCH_ROCM_ARCH under the FROM for that layer, the argument isn't actually in scope and therefore isn't set. This causes flash attention to try to build with the default ...
make _inductor.config.rocm.supported_arch set order deterministic for caching #52258 Sign in to view logs Summary Jobs assign Run details Usage Workflow file Triggered via issue July 26, 2024 19:21 pytorchmergebot commented on #131921 782efd8 Status Success Total duration 14s Artifacts – ...